Prediction Models for Postoperative Delirium Among Cancer Patients: A Scoping Review

Objectives: To systematically map postoperative delirium (POD) prediction models in cancer patients, focusing on study design, modeling methods, performance evaluation, and reporting standards. Methods: Following the Arksey and O'Malley framework and PRISMA-ScR guidelines, nine databases were searched from inception to 24 April 2026. Studies developing or validating POD prediction models in cancer patients were included and narratively synthesized. Results: Thirty-two studies conducted in China, South Korea, the United States, the Netherlands, and Japan were included, most of which had a high risk of bias. POD incidence ranged from 6.70% to 46.39%. The Confusion Assessment Method was the most common assessment tool, and logistic regression was the predominant modeling approach. The most frequently identified predictor domains were age, operation/anesthesia time, preoperative nutritional indicators, preoperative inflammatory indicators, and American Society of Anesthesiologists physical status classification. AUC values ranged from 0.690 to 0.973, and C-index values ranged from 0.783 to 0.963. However, 43.75% of studies only evaluated performance in the development dataset, 25.00% conducted external validation, and 43.75% did not assess clinical utility. Although 87.50% of studies visually presented models, interpretability analyses for machine-learning models were insufficient. Conclusions: Existing POD prediction models for cancer patients show promising discrimination, but their readiness for routine clinical use remains limited by insufficient external validation, inconsistent clinical utility assessment, and incomplete model transparency. These models should currently be regarded as risk stratification tools rather than definitive clinical decision aids. Future studies should prioritize standardized reporting, multicenter external validation, model updating, and decision-curve analysis to support safe clinical translation.